Online machine learning for autonomous earth moving vehicle control

A non-transitory computer-readable storage medium storing executable instructions that, when executed by a hardware processor, cause the hardware processor to perform steps comprising:

• determining, by an earth moving vehicle (EMV), a desired state for a portion of the EMV;
• determining, by the EMV, a current state of the portion of the EMV;
• selecting, by the EMV, a set of control inputs for moving the portion of the EMV from the current state to the desired state using a machine learning model trained to predict control inputs for moving the portion of the EMV to the desired state based on the current state, the set of control inputs comprising a first control input for the portion of the EMV and a second control input for stopping rotation of the portion of the EMV before or when the portion of the EMV reaches the desired state, the second control input based at least in part on a deceleration rate of the portion of the EMV predicted by the machine-learned model;
• executing, by the EMV, the selected set of control inputs;
• determining an updated current state of the portion of the EMV;
• and
• updating the machine learning model based on the updated current state of the portion of the EMV.

A method comprising:

• determining, by an earth moving vehicle (EMV), a desired state for a portion of the EMV;
• determining, by the EMV, a current state of the portion of the EMV;
• selecting, by the EMV, a sequence of control inputs for moving the portion of the EMV from the current state to the desired state, the sequence of control inputs comprising at least a deceleration starting point at which the portion of the EMV will begin slowing down, wherein the sequence of control inputs are selected based on an average of two or more historical control inputs executed by the EMV or one or more additional EMVs;
• and
• executing, by the EMV, the selected sequence of control inputs.